jpt-chat and the AI Writing Assistant Decision: A 5-Step Procurement Checklist

Not long ago, a salesperson asked me why someone in procurement was involved in an AI software decision. Answer: because someone has to deal with the login accounts, the invoices, and the mess when a tool doesn’t do what the demo promised.

I’m an office administrator for a 40-person company. I manage software subscriptions—roughly $60k annually across 15 vendors. And I report to both operations and finance. When we decided to evaluate AI writing assistants, I didn’t care about benchmarks. I cared about whether the tool would fit into how people actually work. After going through this with several platforms (including jpt-chat), I put together a checklist.

This is not a technical deep dive. If you need model architecture, there are better places. This is a buyer’s checklist for getting an AI writing assistant that doesn’t create more work than it saves.

1. Start with your no-go list, not the feature list

Most checklists start with what you want. Mestart with what you don’t want to touch. For us, the no-go list was: no processing of client contracts, no auto-sending emails, no using unapproved vendors. We also decided that AI would not write final client-facing proposals without a human reviewer. This wasn’t because the AI couldn’t do it. It was because accountability can’t be delegated. If a proposal goes out with a wrong number, the account manager has to explain it, not the model.

Once the no-go list was written, everything else got easier. Tools that needed access to our full document drive were off the table. Tools that made it hard to delete data were off the table. Speed, quality, price—you don’t get all three in every scenario. In that order? Actually, for us: safety, speed, price. Pick two.

2. Understand what GPT-4 means and what it doesn’t

People search for “what is GPT-4 and how is it different” because every vendor says they use it. Here’s the non-technical version: GPT-4 is a large language model developed by OpenAI. It’s the engine underneath ChatGPT, and many other products are built on it. It can handle text plus images in some versions, and it can follow longer, more complicated instructions than older models like GPT-3.5. But the model is not the product (i.e., the engine doesn’t control formatting, permissions, or login flow). A tool can be “powered by GPT-4” and still have terrible document formatting, confusing login flows, or weak permission controls.

Per FTC guidance (ftc.gov), claims about AI should be truthful and not misleading. If a vendor can’t explain its AI claims, treat it like any other unsubstantiated marketing claim.

So when a vendor says “GPT-4 powered,” ask what part of the product uses it. Is the chatbot running on GPT-4? Are the templates? Is your private content used to train or fine-tune anything? If they can’t answer clearly, that’s a red flag.

There are plenty of ChatGPT alternative free options, including jpt-chat’s free tier. But free doesn’t mean identical. Check whether the free tier puts your text into a public training pool, whether it limits usage, and whether it allows commercial use. Free is a starting point, not a contract.

3. Check the login and admin experience—not just the consumer login

People often type chat jpt into Google when they mean either ChatGPT or a specific platform like jpt-chat. For an individual user, chat jpt login is just a link. For an administrator, it’s a different story. I almost skipped this step because I was testing as a user. That was a mistake.

When I tried out jpt-chat, I logged in with a standard email link, which was easy. But the questions that mattered for us were not about the consumer login. They were:

• Can we provision accounts in bulk?
• Can we disable an account immediately when someone leaves?
• Is there single sign-on?
• Does the system log admin actions and show who edited what?

If the answer is no, you will either accept the risk or pay someone to build a workaround. I made this mistake years ago with another software purchase. I said “enterprise-grade security.” The salesperson heard “we have SSL.” We discovered the mismatch when our legal team asked for a SOC 2 report and the vendor said none existed. That gap cost us a month of compliance work. Now the admin checklist comes before the feature checklist.

4. Test with your own documents, not their example library

The surest way to evaluate an AI writing assistant is to give it three files you actually use. For us, those were a procurement email, an internal FAQ, and one product description with specifications. We ran each test in the same session. We timed. We checked facts. We looked at tone.

One tool saved us ten minutes on an email but changed a delivery date in the response. Another tool produced clean first drafts but lost our terminology: it wrote “vendor” when the contract said “supplier.” Neither failure showed up in the vendor’s demo.

I’m not an AI researcher, so I can’t tell you why some models handle context differently. What I can tell you from a buyer’s perspective is that the test matters more than the benchmark score. In our tests, jpt-chat handled the tasks cleanly, but I’m not going to claim it’s universally better. It fit our workflow. That’s the only measurement that counts.

I don’t have hard data on how many teams skip this step. Based on the number of conversations I’ve had with colleagues, my sense is most people test with one sample document or none. A single example is not enough. You need to see how the tool behaves with the type of content that will make up 80% of its workload.

Still, the upside was clear: faster drafting, fewer blank pages, more consistent first-pass emails. The risk was inconsistent tone and factual errors. I kept asking myself: is the time savings worth the possibility of a mistake in a client email? For our team, the answer was yes—if and only if a human reviews every output.

5. Calculate the total cost, including the free plan

The sticker price is the least interesting number. Total cost includes setup time, training, admin effort, and the hours your team spends fixing what the AI generated. A free tier that produces output requiring heavy editing is not free. It’s a tax on your team’s patience.

Real talk: we once chose a free plan to save about $50 per month. The tool drafted citations correctly maybe half the time. Our staff spent six hours a month correcting those citations. At every hour worth $40 internally, that’s $240 per month of hidden cost. The paid plan would have been cheaper.

Also check the boring stuff: invoices, payment terms, renewal terms. In 2024, I found a great price from a new vendor—$2,400 cheaper annually than our regular supplier. The vendor couldn’t provide a proper invoice. Finance rejected the expense report. I ended up charging $2,400 to our department budget. Now I verify invoicing capability before placing any order. For SaaS, ask for a W-9, net-30 terms, and a copy of the order form before the trial starts.

A few things I’d do differently next time

If I did this again, I would bring an end user into the trial from day one. I would ask for the data processing agreement before the demo, not after. I would test account deletion first. And I would not assume that a tool being called a “ChatGPT alternative free” means it is a drop-in replacement for anything. Free tools can be genuinely useful, but they are not all the same.

The uncomfortable truth is that no purchase comes with a guarantee. You will make a call with incomplete information, and that’s okay as long as you know what you didn’t check. One informed question is worth ten sales demos. The point of this checklist isn’t to make you an AI expert. It’s to help you ask the questions that matter for your team.